Vals Index v2.1 adds Terminal Bench 4 and a tax agent category

Version 2.1 of the Vals Index changes what the index measures, so a model's score under it is not a clean continuation of its earlier score. Vals AI, which runs the index, announced two changes. It swapped Terminal Bench 2.1, one of the benchmarks inside the index, for its newer successor, Terminal Bench 4. It also added its own Tax Agent Benchmark as a separate category, which Vals describes as "weighted by GDP." Vals calls the tax benchmark proprietary, meaning Vals built and holds the test itself rather than adopting a public one.

An index like this combines several benchmark results into one number, so changing a component or adding a category changes the number even if no model changes. If a model rises or falls across the version boundary, part of the move can come from the new tests rather than from the model. Rankings published before the update, such as Opus 5.5 taking first place in late September, were measured on the earlier mix. The index has been revised again since: Vals reported Gemini 4 Carbon's debut on v2.2 in late October. The post does not give new scores or say how the GDP weighting is calculated.